Subagents and coding agents

Some work does not belong in the main conversation — a wide codebase search that would flood the context window, or a multi-file refactor that deserves its own agent. The delegateTask tool hands that work off and brings back the result.

There are two delegation backends, and you pick one globally.

Worker mode (default)

Yachiyo runs the subagent itself, using your configured provider. Four named agents, each with a fixed role:

Agent	Use it for
Explore	Finding files, searching patterns, understanding how something works. Read-only. Run several in parallel when the modules are independent.
Plan	A step-by-step strategy for a complex multi-file change, with the critical files identified.
Review	A second opinion on uncommitted changes before you commit.
General	Editing, running commands, anything the read-only agents cannot do. The fallback.

All four are enabled by default. Which ones are available is set in Settings → Capabilities → Coding agents.

Subagents start blank

A worker subagent sees only its system prompt and the prompt it is handed — not your conversation, not the plan document, not earlier tool results. The task brief has to be self-contained, with explicit file paths and done criteria.

This is why “continue with the plan above” fails and “implement the retry logic in src/net/client.ts:120, keeping the existing backoff constants” works.

Each worker can run on its own model — Settings → Capabilities → Coding agents has a model selector per worker, defaulting to whatever model is making the call. Read-heavy Explore work on a cheap model while the main conversation stays on a strong one is the obvious use.

ACP mode (deprecated)

The other backend hands the task to an external coding agent over the Agent Client Protocol — Claude Code, Codex, or anything else speaking ACP. The agent runs as its own process with its own model and context, then reports back into the thread.

Deprecated

ACP mode is deprecated. Existing profiles keep working for compatibility, but worker mode is where the development is. Do not build new setups on it.

One profile ships preconfigured for it:

Field	Value
name	Claude Code
command	npx
args	["-y", "@zed-industries/claude-agent-acp"]
env	{ "ACP_PERMISSION_MODE": "acceptEdits" }

In ACP mode, delegateTask is unavailable unless the workspace is a git repository. An external agent editing files with no way to diff or revert is not a trade worth making, so the tool refuses rather than degrading quietly.

Switching modes and managing profiles

yachiyo agent mode worker      # or: acp
yachiyo agent list             # current mode, plus any ACP profiles

ACP profiles take these fields:

Field	Type	Notes
id	string	Auto-generated on add
name	string	Display name
enabled	boolean	Disabled profiles stay in config but are not offered
description	string	Shown in the UI
command	string	Executable — npx, node, an absolute path
args	string[]	Arguments
env	Record<string,string>	Extra environment variables

Full details in the agent CLI reference.

Or skip delegation

Delegation is a judgement call, not a setting. For small, well-scoped changes the main agent doing the work itself is often faster than writing a task brief good enough for an agent that cannot see your conversation — so just ask for the change directly. If you want the boundary enforced rather than implied, Explore and Chat run modes drop delegateTask from the tool list entirely.